Deep learning model observer for 4-alternative forced choice in digital breast tomosynthesis

Abstract

The purpose of this study is to investigate deep learning model observer (DLMO) using classification network for various sizes of mass detection tasks in digital breast tomosynthesis (DBT). Different tube current and sweep angular range acquisition settings were repeated to acquire different DBT images. As a result, a total number of 66 cases of four alternative forced choice (4AFC) reading was undertaken to the human observer and DLMO. The images of spheroidal mass with different sizes of 1.8, 2.3, 3.1, 3.9, 4.7, and 6.3 mm in the target slab of CIRS breast phantom were cropped by 200x200 size for their region of interest (ROI). A percentage of correct responses (𝑃c) was measured at the end of each human observer test and compared by the accuracy of prediction using DLMO. The results indicated that our proposed DLMO showed the averaged training accuracy 93% using 60 testing datasets after 50 epochs of training using 204 input training datasets. Both the accuracy and loss function in training session of categorical cross-entropy reached a plateau after 35 epochs. The sensitivity and specificity of the testing data from DLMO showed 78% and 98%, respectively, which is comparable to the 𝑃c, 0.89, on 4AFC test from the human observer. Although the current number of datasets is too small to apply in clinical trials, the authors considered that our DLMO on the phantom study is useful for initial trial of analysis on DBT images.